Review for NeurIPS paper: Assessing SATNet's Ability to Solve the Symbol Grounding Problem

Summary and Contributions: EDIT after author response: I thank the authors for their thoughtful response, and am happy to see the authors will address the issues around the framing/tone of the paper. I also like the suggestion to include a discussion of the different definitions of end-to-end; this discussion is likely to be of great value to the community, given the overloaded definition of end-to-end. To address these claims/elaborate further -- * SATNet-specific: In cases where the logical rules aren't known, it's not possible to use a discrete solving module, since the rules need to be learned. That is, using a classifier and a separate Sudoku solver wouldn't work for the SATNet experiments, as the premise of the SATNet experiment was that the rules of Sudoku weren't known. This allows learning to, e.g., be more data-efficient.